Fix bug #307 : wrong NIC selection on the reduction tree.
The reduction tree (tree up) was inverting the NICs to use, causing performance issue in cases where we are using different NICs on a given channel.
Этот коммит содержится в:
@@ -27,7 +27,7 @@ void ncclTopoFree(struct ncclTopoSystem* system);
|
||||
ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm* comm);
|
||||
|
||||
// Query topology
|
||||
ncclResult_t ncclTopoGetNetDev(struct ncclTopoGraph* graph, int dir, int channelId, int* net);
|
||||
ncclResult_t ncclTopoGetNetDev(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, int rank, int channelId, int* net);
|
||||
ncclResult_t ncclTopoCheckP2p(struct ncclTopoSystem* system, int64_t id1, int64_t id2, int* p2p);
|
||||
ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* topo, int64_t busId, int netDev, int read, int* useGdr);
|
||||
|
||||
|
||||
Ссылка в новой задаче
Block a user